A ReAct agent is an LLM-driven system that alternates between reasoning and acting—often called reason + act or ReAct prompting—so it can decide what to do next, use tools, and update its plan from the results.
ReAct is useful when a model should not just produce an answer, but also take steps in the world: search a document set, call an API, inspect a database, or ask for more information.
In practice, teams reach for it when:
It is a simple and durable pattern for building agentic behavior without requiring a full autonomous planner.
The core idea is to interleave two kinds of output:
After the action returns, the model reads the observation and repeats the loop until it can answer. This makes the process more grounded than “think once, answer once,” because each step can be informed by fresh evidence.
In the original ReAct paper, this reasoning-and-acting pattern was presented as a way to improve task performance on tasks that benefit from both language reasoning and environment interaction. In practice, “ReAct agent” usually means an agent implementation that follows that pattern, even if the exact prompt format varies.
User: “What is the capital of the country whose largest city is Sydney?”
A ReAct-style agent might do:
The important part is not the exact wording—it’s the loop: think, act, observe, repeat.
